KR101183391B1 - Image comparison by metric embeddings - Google Patents

Abstract

A system and method for image comparison by metric embedding is presented. In one implementation, graphs are generated and compared from each image. The graph metric is then embedded in the groups of trees for each image. The minimum differences between each of the groups of trees for different images are compiled into a matrix from which a similarity measure for image comparison is obtained.

Metric embedding, similarity measure, shape parameter

Description

Image comparison by metric embedding {IMAGE COMPARISON BY METRIC EMBEDDINGS}

To search for text, images, ideas, websites, etc. on the Internet or a computing device, a text keyword can be entered to initiate the search. Background Art Image retrieval in a computing environment is conventionally limited to the input of text keywords summarizing the image or the text summarizing visual objects or visual features of the image. Conventionally, it has been difficult to search an image without converting some of the key parts of the search into text. Thus, a conventional image search is actually a text search.

The reason why a successful image search is difficult is in the image itself. Images are much more complicated than words. This is because the image can be modified so that the modified image is mathematically different from the original image, but both the modified image and the original image are still indistinguishable from human vision. Thus, if the user brightens or darkens the image, changes the contrast, fine tunes the color, crops, or compresses the image, the parameters that quantify the image change considerably, even though the image may look the same. have. To avoid this complexity, it was easier to summarize a picture into words and then perform a word search on the picture.

Image retrieval can be initiated by entering the image itself, and mathematically and informatively, so that the image can be successfully completed in many different ways, for example, by finding a matching version of the modified image using image editing software. We need a way to capture information informally.

We present a system and method for comparing images by metric embedding in a form that allows digital comparisons without directly relying on the visual appearance of the images being compared. In one embodiment, a graph, such as an undirected graph with vertices and edges, is generated from each image for comparison. Various features of the pixels can be used to generate a graph. Generating a graph from an image can be accomplished in part by a wavelet transform technique that obtains coefficient vector values. You can add a noise reduction process to create a reliable graph for comparison. For example, after the wavelet transformation process, you can zero out the meaningless values and downscale the image or graph.

The graph metric is then embedded in trees, such as a group of hierarchical well-separated trees (HST) for each image. The HST for an image may be circular clustered until the graph representing the image is split.

Groups of trees with embedded graph information are distinguished to obtain a similarity measure for the images they represent. In one implementation, because the groups of trees representing the image are cyclically generated, the tree discrimination technique is repeatedly applied to the root HSTs and their associated child trees. In one implementation, the differentiation comprises compiling a matrix of minimum differences between the images and vectorizing the matrix so that the normalization process averages or centers the summary of the minimum differences between the images in a single similarity measure. (median) can be found.

1 is an exemplary diagram of a graph representing an image.

2 is an exemplary diagram comparing an image by distinguishing groups of trees in which a graph metric is embedded.

3 is an exemplary diagram of an image comparison system.

4 is an exemplary block diagram of an image comparison engine.

5 is an exemplary diagram of image comparison by iterative comparison of sets of child nodes.

6 is an exemplary flow chart of a method for image comparison.

summary

The system and method described herein provide image comparison techniques. These image comparison techniques can be used to perform an internet image search or to identify a modified version of the source image within a pool of images or the like. For example, a user may modify the source image by photo editing software or by cropping and compressing the source image to include the image in an email message. The systems and methods described in this specification provide a way to search for and find these modified images among other images.

1 is an example implementation of an image comparison method in which pixel values of an image 100 are represented by an undirected graph 102. Each vertex in the undirected graph 102 represents a corresponding pixel in the image. In one implementation, the image preprocessing step of generating the graph 102 includes the application of a wavelet transform to set coefficient vector values, and the dropping of very small values (less than the selected threshold). The edges between the vertices in the undirected graph 102 may be weighted according to the difference in values between the pair of pixels at both ends of the edge. The pixel feature used to determine the weight of the edges can be selected from among many different types of pixel features such as density, energy, and the like. The non-directed graph 102 thus generated may be downscaled to save processing power and reduce noise. Such an undirected graph 102 may be referred to roughly as a preprocessed image herein because the further processing follows the graph generation process.

The group 104 of trees, such as HST with embedded graph metrics, can be derived from the undirected graph 102 generated in the image preprocessing step described above. The group of trees 104 is set by a clustering method that sufficiently repeats the initial cluster to partition the preprocessed image into well-separated trees. A typical group of trees 104 for this method includes approximately ten (parent and child) trees.

As shown in FIG. 2, the differentiation technique includes a group of trees 104 for the source image and each group 204 for a number of other images to be compared to find one image among a group of images. 204 ', 204 ", 204'"). The image with the best similarity measure for the source image 100 is selected with the best match.

In one implementation, the discrimination technique employed employs the effect by assigning shape vectors to each node in each tree and employing similarity measures that result in a minimum difference between corresponding nodes for the different images being compared. Exert. That is, a matrix of differences is aggregated across the trees in two groups of trees representing two images. This matrix is assigned a vector and its components are averaged or otherwise normalized to obtain a similarity score. Thus, to find a copy of one image or a modified version of that image in a pool of images, the similarity scores can be compared to reveal the matchability.

Example system

3 shows an example system 300 for comparing images. Computing device 302 is communicatively coupled to the Internet 304. Computing device 302 hosts image comparison engine 306. This type of networked computing environment is shown by way of example only. Other types of computing and network environments can also host targets. Image comparison engine 306 enables image search 308. The target or source image 310 can be selected by a user or by a component of the system 300. The image comparison engine 306 compares the source image 310 with other images found on the Internet 304, for example, and returns a similarity measure for each comparison. The image comparison engine 306 may select the best match between the source image 310 and the members of the set of candidate images.

Instead of retrieving a set of images from the Internet 304, the image comparison engine 306 may retrieve a set of images, such as in a file directory or database, or stored in a folder as an attachment to an email message. .

Example engine

4 illustrates the example image comparison engine 306 of FIG. 3 in more detail. The image comparison engine 306 of FIG. 4 is presented as an example of configuration for overview. Many other configurations of the illustrated components or similar components are possible. This image comparison engine 306 may be implemented in hardware, software, or a combination of hardware and software.

An exemplary configuration of the image comparison engine 306 includes an image graphing engine 402 and an image comparator 404. Image graphing engine 402 generates graphs and groups of trees 104 representing images, while image comparator 404 is between groups of trees 104 to match the candidate image with source image 100. Find the similarity measure.

The image graph engine 402 further includes a grascaler 406, wavelet transform engine 408, and downscaler 410, which are image preprocessing components. These components generate a graph that represents a preprocessed image 412, ie, an image 100 and a group of trees 104 that allow for comparison of the images. Image graphing engine 402 further includes a tree generation engine 414. The tree generation engine 414 further includes a cluster recurser 416, a tree pruner 418, and a redo filter 420. They create groups of various trees (eg, 104, 104 ') with embedded graph metrics to represent different images.

The image comparator 404 introduced above further includes a tree discrimination engine 422 and an aggregator 424. The tree discrimination engine 422 further includes a minimum difference engine 426 that further includes a tree topology comparator 428 and a cluster topology comparator 430. Aggregator 424 further includes a matrix 432 of differences as well as a normalizer 436 and a vector engine 434 that produce a similarity measure 438 that quantifies the similarity between a pair of images.

Image graphing engine

Among the image preprocessing components of the image graphing engine 402, the grayscaler 406 may receive a color RGB image input and convert this input to grayscale, for example by an L2 norm of RGB values. Can be.

In one implementation, wavelet transform engine 408 performs three-level Haar wavelet transform. As a result, small values in the coefficient vectors can be thresholded to zero, that is, coefficients whose absolute value is smaller than the threshold (eg 0.15) are changed to zero. The wavelet reconstruction procedure is then applied to these coefficient vectors, and downscaler 410 downscales the results to about 70 pixels in width to generate preprocessed image 412.

Tree generation engine 414 generates k HSTs from preprocessed image 412. Typically, k = 10. Such HST may be generated by cyclic hierarchical clustering decomposition. The tree generation engine 414 starts by setting the initial cluster of graph node C ₀ equal to the preprocessed image 412. The cluster size parameter is then designated r and the number of pixels in preprocessed image 412 is designated N. Typical values of r range from about 1 to about −2. You can improve the result by making some adjustments to the r parameter.

The cluster recurser 416 selects the cluster C _i and selects a random point x∈C _i . If all points of C _i are within distance rlog (N) from x, no further processing is needed and C _i is returned as a tree. If not, the radius d is selected from the inverse-type distribution. That is, when the preprocessed image 412 includes N pixels, for a predetermined value of the parameter r, the cluster recurser 416 selects the value of x evenly at [0,1] and d = -r ( set to log N) (log x).

The cluster recurser 416 then sets C _{i + 1} = B (x, d) so that the ball is centered on the new x of radius d from the previous x, and the tree is rotated at C _{i + 1} . Set C '= C _i / C _{i + 1} . This procedure is repeated replacing C _i with C '. Therefore, to form the thus formed circular tree C _i C _i is the child until the split. As a result, the trees become a group 104 of HST trees in which the metrics of the preprocessed image 412 are embedded.

The tree pruner 418 reduces the effect of noise in image comparison by pruning nodes whose corresponding clusters contain fewer than a predetermined threshold pixel number (typically 100). Note that C _{{i + 1}} = B (x, d) ∩C _i , that is, the cluster is not the same as the ball but the set of pixels contained within the ball. The set of pixels generally does not have a very ball-like shape, so the shape vectors described below are additionally important. In one implementation, tree pruner 418 can eliminate the initial step of ensuring that the size of the cluster is at most r (logN) by pruning small nodes at the end of tree generation.

Similarly, redo filter 420 reduces noise by rerunning the entire tree generation process if the final tree has fewer nodes, eg, less than six, nodes. In other words, since a small tree indicates poor clustering, this tree must be regenerated with different clustering parameters.

Image comparator

In the example of the illustrated image comparator 404, the tree discrimination engine 422 receives a tree, such as HST, from the group 104 of trees for the source image 100. Similarly, tree discrimination engine 422 receives a tree from group 104 'of trees for the candidate image.

를 각 노드 v에게 지정한다. 이 파라미터들은 v에서의 클러스터의 컨벡스 헐(convex-hull; H)을 컴퓨팅하고 d_max를 H에서의 임의의 2개 점들 간의 최대 거리로 설정함으로써 구해지며, 또한, d_min = min_x∈H max_y∈Hρ(x,y)이며, 여기서 ρ(x,y)는 이미지 그래프에서 x에서부터 y까지의 거리이다.Given these two trees (e.g., HST), the tree discrimination engine 422 has a minimum difference engine 426 that produces a difference measure between these trees. In one implementation, tree topology comparator 428 considers the topology of the trees while cluster topology comparator 430 considers the topology of the clusters associated with the tree nodes. Given a child HST, the minimum difference engine 426 scales the weight of each edge to, for example, the diameter of the parent. The minimum difference engine 426 concatenates the position of the cluster at each node v by the shape parameters d _min , d _max

Is assigned to each node v. These parameters are obtained by computing the convex-hull (H) of the cluster at v and setting d _max to the maximum distance between any two points in H, and also d _min = min _x∈H max _y∈H ρ (x, y), where ρ (x, y) is the distance from x to y in the image graph.

및

로 설정된다. 트리 구별 엔진(422)은

로 설정하고, 여기서, a_v는 전술한

에서의 형상 벡터이다. 이후, 최소 차이 엔진(426)은 s₁을 δ_ij를 최소화하는 쌍(i,j)으로 하고, s₂를 s₁의 행과 열이 제거된 δ_ij를 최소화하는 쌍으로 한다. s_i의 이러한 시퀀스는 T(502)의 루트의 자식들과 U(502')의 루트의 자식들 간의 매칭을 설명한다. k_T ≠ k_U이면, 수 개의 자식 트리들이 매칭되지 않을 것이다.As shown in FIG. 5, with two HST T 502 and U 502 ′, the tree discrimination engine 422 cyclically computes the tree discrimination function " treediff (T, U) " Mapping is done between sets of nodes. The children of the roots of T 502 and U 502 'are each

And

Is set to. The tree discrimination engine 422

, Where a _v is the above

Shape vector at. The minimum difference engine 426 then sets s ₁ as a pair (i, j) to minimize δ _ij and s ₂ as a pair to minimize δ _ij from which the rows and columns of s ₁ have been removed. This sequence of s _i describes the matching between the children of the root of T 502 and the children of the root of U 502 '. If k _T ≠ k _U , several child trees will not match.

The tree discrimination engine 422 can match the matched children (eg, top level matched children 504, 504 '), 506, 506', 508, 508 ', and 510, 510'. Cyclically compute the treediff function for its lower-level matched children), such as), producing d _i from the pair s _i . To do Σd _i , the tree discrimination engine 422 adds the difference between each pair of edge weights, and the difference of a _v vectors. If there are any unmatched children of T or U, the tree discrimination engine 422 adds the total edge weights of these subtrees (but does not use a _v vectors). The tree discrimination engine 422 obtains the final total value, which is the treediff function value for T and U.

In one implementation, in addition to the children matched by the minimum value of the shape vector, the tree discrimination engine 422 also performs several levels of exhaustive search. In some situations, this may provide better results. If treediff (T, U, 0) is the distinguishing function described above, treediff (T, U, d) focuses on producing all possible matches between the children of T and U. Thus, from the highest level, the tree discrimination engine 422 calls treediff (T, U, d-1) for each pair in each match and returns the mathematical minimum of the total value of the match for every match. The number of children of T and U is not large because the tree pruner 418 described above generally optimizes the trees by pruning nodes whose corresponding clusters contain fewer than the predetermined threshold number of pixels. It is possible to create a thorough match because the typical values of d in the top level call to treediff (T, U, d) are only 2 or 3.

In the context of comparing two images (eg, G and H) to obtain a similarity measure, the image graphing engine 402 described above computes k HSTs from each image (typically k = 10). The groups of trees 104 and 104 'for each image are {T _i } and {U _i }, respectively, where 1 ≦ _i ≦ k. Aggregator 424 forms Δ _ij = treediff (T _i , U _j ) as matrix 432 of differences. The vector engine 434 forms a vector d by d _i = min _j Δ _ij . In some implementations, this minimum operation is not symmetrical, thus Δ _ij ≦ Δ _ji . The normalizer 436 then aggregates the results of the treediff function, such as Δ (G, H) = mean (d) or Δ (G, H) = median (d). The latter case, which is an aggregate by median, may yield better results in some embodiments, because it is less sensitive to outlier values (as with the averaging filter versus the median filter used for noise images). Because.

In another example, an edge detector (not shown) is used for clustering in the tree generation engine 414. Thus, after the low pass effect of the wavelet transform engine 408, the images can be processed by a Canny edge filter (not shown), so that the trees can be clustered. Because the Canny edge filter does not guarantee that objects are surrounded by a closed edge curve, the image graph used for HST in this case can be generated using Gaussian averaging to compute pixel value differences. This has the effect of preventing clustering from crossing small gaps between corner endpoints. If the tree discrimination engine 422 applies a Canny edge filter in this way, the final HST can in many cases yield good results.

Example method

6 shows an example method 600 for image comparison. In the flowchart, the operations are summarized in separate blocks. The example method 600 may be performed in hardware, software, or a combination thereof, for example, by components of the example image comparison engine 306.

In block 602, a graph is generated from each image to be compared. Graphing an image so that the image can be compared digitally for similarity or dissimilarity with other images constitutes the image, for example, by weighting graph edges with the difference values of adjacent pixels in the image. Quantitatively quantifying pixels. If the image is color, one can create a graph using the color components of the pixels, or can grayscale a color image and create a graph using various other non-color features of the pixels. For example, the example image comparison engine 306, which is an engine that generates a graph from an image, may include other preprocessing components that make the graph a reliable indicator of the image for comparison.

Generating a graph from an image can be facilitated by a wavelet transform to obtain coefficient vector values. One of the key factors in creating a reliable graph to compare images is to remove extraneous noise from the graph. Thus, after wavelet transformation, it is possible to zero the meaningless values and thus reduce the overall amount of information that this method must process. Generating graphs can also benefit from downscaling images and / or graphs to efficient sizes.

In block 604, a group of trees is generated for each graph. In this part of the example method 600, the graphs representing the images while quantifying the images to be compared are still not suitable for mathematical comparison techniques that can be executed digitally. Thus, rather than comparing graphs directly, the metrics of each graph are embedded into a group of trees. For example, HST is suitable for mathematical comparison techniques that can be implemented digitally efficiently. Thus, the image is graphed, which in turn is a group of related trees. By clusters and nodes of the trees, the images can be compared mathematically. That is, it can be manipulated digitally and compared in a form that does not directly depend on the visual quality of the images for comparison.

In block 606, the groups of trees for the two images to be compared are distinguished from each other to obtain a similarity measure. In other words, the groups of trees with embedded graph information are compared to obtain a similarity measure for the images they represent. In one implementation, the tree discrimination technique is repeatedly applied to the root HSTs and their associated child trees because the group of trees representing the image is cyclically generated. In one implementation, the discrimination includes compiling a matrix of differences and vectorizing the matrix, such that the averaging process (ie, central find) can summarize the differences into one similarity measure.

The system and method described herein can embed graph metrics into tree metrics with approximately low distortion. Given a general graph defining the metric, embedding into the tree metric has a Ω (n) distortion, where n is the number of points in G. However, with G → T _i embedding, with a group of tree metrics {T _i } and a graph G (with two random points (x, y∈G)), a random selection of T _i The expected distortion of x, y → T _i with respect to is O (log n).

conclusion

The above-described subject matter of the present invention may be implemented in hardware, software, firmware, or the like, or a combination thereof. In certain implementations, the subject matter of the present invention may be described in the general context of computer-executable instructions, such as program modules, being executed by a computing device or communication device. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The subject matter of the present invention may also be practiced in distributed communications environments where tasks are performed via wireless communications by remote processing devices that are linked through a communications network. In a wireless network, program modules may be located in both local communication device storage media and remote communication device storage media, including memory storage devices.

The foregoing description describes example systems and methods for image comparison. Although the subject matter of the invention has been described in language specific to structural features and / or methodological actions, it will be understood that the subject matter defined in the claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are disclosed as examples of forms of implementing the claims.

Claims (20)

A computer-executable method executable on a tangible computer and assisting in comparing images, In the computer, representing a first image as a first graph and a second image as a second graph, wherein each vertex in the first graph and the second graph corresponds to a pixel in each image -; At the computer, assigning weights to each corner between vertices in each undirected graph, the weights corresponding to differences between adjacent pixel values; And In the computer, deriving a first family of trees from the first graph and a second group of trees from the second graph, embedding graph metrics in the groups of trees. Contains-, Deriving a group of individual trees from the computer includes deriving hierarchical well-separated trees by recursive hierarchical clustering decomposition for the image. , The cyclic hierarchical clustering decomposition, In the computer, selecting a cluster of pixels from the image, a random pixel location x in the cluster of pixels, a cluster size r and a radius d equal to -r (log N) (log x), where N is the image Is the number of pixels in-, In the computer, generating child clusters of the cluster of pixels until the cluster of pixels is partitioned, the generating recursively contstructing child clusters at successive radii from x. Each child cluster comprises a previously obtained cluster set to a ratio divided by the cluster at the current radius, The method also includes Determining, at the computer, a difference measure for the first image and a second image based on a difference measure between the first group of trees and the second group of trees; And Displaying, at the computer, the determined difference measure And a computer-executable method to assist in comparing the images. The method of claim 1, Representing the first image as a first graph and the second image as a second graph comprises processing each image with a wavelet transform, the computer-implementing aid of the comparison of the images. Possible way. 3. The method of claim 2, Wherein the wavelet transform comprises a three-level Haar wavelet transform. 3. The method of claim 2, In the computer, further downscaling each image to reduce noise to produce a preprocessed image. 3. The method of claim 2, In the computer, further downscaling each image to 70 pixels in width to form a preprocessed image. The method of claim 1, Determining a difference between the first group of trees and the second group of trees comprises comparing a topology of the first group of trees and the second group of trees and the tree of the first group and the second group A computer-executable method to assist in comparing images, comprising comparing the topology of clusters related to nodes. The method of claim 1, Determining the difference between the first group of trees and the second group of trees, Computing a convex hull of a cluster associated with each node of the first group of trees and the second group of trees to obtain shape parameters for each node, Assigning a shape vector to each node by concatenating the position of the cluster with respect to the node with the shape parameters; and Finding a difference measure that minimizes the difference between the shape vectors for each of the sets of child nodes of the first group of trees and the second group of trees And a computer-executable method to assist in comparing the images. The method of claim 7, wherein Aggregating, in the computer, the differences of the shape vectors across the nodes of the trees by forming a vector of minimized differences of the shape vectors and normalizing the components of the vector. Computer-executable method to assist the comparison of the data. The method of claim 1, In the computer, further comprising repeating the method continuously with respect to the first image and the members of the set of images to find a member of the set having the best similarity to the first image. Computer-executable method to assist. One or more tangible computer readable storage media storing computer-executable instructions that when executed by a computer assist in the comparison of images. In the computer, representing a first image as a first graph and a second image as a second graph, each vertex in the first graph and the second graph corresponding to a pixel in each image; In the computer, assigning a weighting value to each corner between vertices in each undirected graph, the weighting value corresponding to a difference between adjacent pixel values; And In the computer, deriving a first group of trees from the first graph and a second group of trees from the second graph, wherein the graph metrics are embedded in the groups of trees; Deriving a group of respective trees in the computer includes deriving hierarchical separated trees by circular hierarchical clustering decomposition for the image, The cyclic hierarchical clustering decomposition, In the computer, selecting a cluster of pixels from the image, a random pixel position x in the cluster of pixels, a cluster size r, and a radius d equal to -r (log N) (log x)-N being the Is the number of pixels in the image In the computer, generating child clusters of the cluster of pixels until the cluster of pixels is divided—the generating comprises circulating child clusters at successive radii from x, each child Cluster includes the previously obtained cluster divided by the cluster at the current radius, The method also includes Determining, at the computer, a difference measure for the first image and a second image based on a difference measure between the first group of trees and the second group of trees; And Displaying, at the computer, the determined difference measure One or more types of computer readable storage media comprising. The method of claim 10, And displaying the first image as a first graph and the second image as a second graph comprise processing each image with a wavelet transform. 12. The method of claim 11, And the wavelet transform comprises a three-level down wavelet transform. 12. The method of claim 11, The method further comprises: at the computer, downscaling each image to reduce noise to produce a preprocessed image. 12. The method of claim 11, And the method further comprises downscaling each image to 70 pixels in width to form a preprocessed image in the computer. The method of claim 10, Determining a difference between the first group of trees and the second group of trees comprises comparing a topology of the first group of trees and the second group of trees and the tree of the first group and the second group One or more types of computer readable storage media comprising comparing topologies of clusters related to nodes. The method of claim 10, Determining the difference between the first group of trees and the second group of trees, Computing the convex hull of a cluster associated with each node of the first group of trees and the second group of trees to obtain shape parameters for each node, Assigning a shape vector to each node by linking the position of the cluster with respect to the node with the shape parameters; and Finding a difference measure that minimizes the difference between the shape vectors for each of the sets of child nodes of the first group of trees and the second group of trees One or more types of computer readable storage media comprising. 17. The method of claim 16, The method further comprises a step of aggregating, at the computer, the differences of the shape vectors across the nodes of the trees by forming a vector of minimized differences of the shape vectors and normalizing the components of the vector: One or more types of computer readable storage media. 17. The method of claim 16, The method further includes, in the computer, continuously repeating the method on the first image and the members of the set of images to find a member of the set having the highest similarity with respect to the first image. One or more types of computer readable storage media. delete delete

Images